How to Prompt AI to Use Your Own Terminology

Your product has Workspaces. The model writes "projects". You correct it, it complies for two paragraphs, then writes "projects" again. To prompt AI to use your own terminology, give it a glossary that names the wrong words as well as the right ones, and put it at the end of the prompt rather than the start.

The one-directional glossary is where almost everyone starts, and it is why this keeps failing.

Why saying the right word is not enough

"Use the term Workspace" tells the model your preference. It does not tell it that "project", "folder", "board", and "team space" are the words it will otherwise reach for.

That matters because the substitution is not random. The model is pulled toward whatever term is commonest across everything it has read, and for most products that is a generic industry word. Your term is the rare one. Every sentence is a small contest between your instruction and a much larger statistical prior, and across a long output the prior wins often enough to be noticeable.

Naming the competitors changes the instruction from a preference into a constraint with a checkable condition.

How to prompt AI to use your own terminology: the glossary block

TERMINOLOGY. These rules override your defaults. Follow them in every sentence.

| Use | Never use | Note |
|---|---|---|
| Workspace | project, folder, board, team space | Always capitalised |
| member | user, seat, account | A person in a Workspace |
| run | job, task, execution | One execution of a flow |
| flow | workflow, pipeline, automation | The thing a run runs |

If you need a concept not in this table, use plain language and flag it in
square brackets rather than inventing a product term.

Three details carry the weight. The Never column is the substitution list and it is the part that works. The Note column prevents the correct word being used for the wrong concept, which is a subtler failure than the wrong word. The final clause stops invention, which is otherwise what happens when the model needs a word for something your glossary did not anticipate.

Put it last

Position matters more than people expect. A glossary at the top of a long prompt competes with everything that follows it; the same glossary immediately before the task is the last thing read and holds better across a long output.

For genuinely long generations, repeat the four or five terms that matter most right before the task, even if the full table is above. Repetition in a prompt is not a sign of a badly written prompt. It is positional reinforcement, and it is cheap. Related positional effects are covered in how to prompt AI to follow your style guide every time.

Terminology is not style

These are different problems and conflating them makes both worse.

The last row is the practical difference. Terminology compliance is mechanically checkable, which means you can verify it without rereading anything. Keep the banned list somewhere your review step can grep for.

Verify, do not trust

Ask for the check as part of the output:

After the main output, add a section titled TERM CHECK listing:
- any banned word you used, and where
- any concept you had to describe without a glossary term

If you used no banned words, write "TERM CHECK: clean".

This is self-reporting, so it catches most but not all. Treat it as a first pass over a long document rather than a guarantee, and spot-check the actual text. A model that has just written "project" four times will sometimes report clean.

Keep the glossary somewhere real

The glossary becomes genuinely valuable once it lives outside any single prompt, because the same block serves documentation, support replies, and marketing copy. Maintain it as a file, paste it in, and update it in one place when the product renames something.

Terminology is one layer of the larger problem of giving a model the facts about your business it cannot infer; how to give AI context about your business covers the rest, and what is context engineering is the wider frame. For storing the glossary alongside the prompts that use it, keeping prompts somewhere shared applies, and when a term changes, tracking that change across your prompts is what stops the old word reappearing in six places. The constraint-first habit underneath all of this is prompt engineering.

FAQ

Why does AI keep using the wrong word after I correct it?

Because the generic industry term is far commoner in training data than your term. Correcting once sets a preference; listing the banned alternatives sets a constraint, which holds better.

Where should the glossary go in a long prompt?

At the end, immediately before the task. For long outputs, repeat the few most important terms there even if the full table appears earlier.

How many terms should a prompt glossary contain?

Enough to cover the terms that actually get substituted, usually five to fifteen. A sixty-term glossary dilutes attention and is better handled by review than by prompt.

Can I use a glossary to enforce tone as well?

Not effectively. Terminology has one correct answer per concept and is checkable by search; tone is a range and needs examples instead.